Pablo Marquez Tello | 6777359 | 2023-10-06 13:49:44 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021-2023 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | |
| 25 | #ifndef ACL_SRC_CPU_KERNELS_SUB_NEON_IMPL_H |
| 26 | #define ACL_SRC_CPU_KERNELS_SUB_NEON_IMPL_H |
| 27 | |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/core/utils/misc/Traits.h" |
| 32 | |
| 33 | #include "src/core/helpers/WindowHelpers.h" |
| 34 | #include "src/core/NEON/wrapper/intrinsics/intrinsics.h" |
| 35 | #include "src/core/NEON/wrapper/scalar/sub.h" |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | namespace cpu |
| 40 | { |
| 41 | template <typename T> |
| 42 | void sub_same_neon( |
| 43 | const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) |
| 44 | { |
| 45 | /** SIMD vector tag type. */ |
| 46 | using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>; |
| 47 | |
| 48 | bool is_sat = policy == ConvertPolicy::SATURATE; |
| 49 | |
| 50 | // Create input windows |
| 51 | Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()); |
| 52 | Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| 53 | |
| 54 | // Clear X Dimension on execution window as we handle manually |
| 55 | Window win = window; |
| 56 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 57 | |
| 58 | constexpr int window_step_x = 16 / sizeof(T); |
| 59 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 60 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 61 | const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x(); |
| 62 | |
| 63 | Iterator input1(src0, window.broadcast_if_dimension_le_one(src0->info()->tensor_shape())); |
| 64 | Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape())); |
| 65 | Iterator output(dst, window); |
| 66 | |
| 67 | if (is_broadcast_across_x) |
| 68 | { |
| 69 | const bool is_broadcast_input_2 = input2_win.x().step() == 0; |
| 70 | Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; |
| 71 | Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; |
| 72 | const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0; |
| 73 | const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0; |
| 74 | |
| 75 | // Clear X Dimension on execution window as we handle manually |
| 76 | non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 77 | |
| 78 | Iterator broadcast_input(broadcast_tensor, broadcast_win); |
| 79 | Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); |
| 80 | Iterator output(dst, win); |
| 81 | |
| 82 | execute_window_loop( |
| 83 | win, |
| 84 | [&](const Coordinates &) |
| 85 | { |
| 86 | const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr()); |
| 87 | const auto output_ptr = reinterpret_cast<T *>(output.ptr()); |
| 88 | |
| 89 | const T broadcast_value = *reinterpret_cast<const T *>(broadcast_input.ptr()); |
| 90 | const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); |
| 91 | |
| 92 | // Compute S elements per iteration |
| 93 | int x = window_start_x; |
| 94 | for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| 95 | { |
| 96 | const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x); |
| 97 | auto res = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v) |
| 98 | : wrapper::vsub(broadcast_value_vec, non_broadcast_v); |
| 99 | if (is_broadcast_input_2) |
| 100 | { |
| 101 | res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{})); |
| 102 | } |
| 103 | wrapper::vstore(output_ptr + x, res); |
| 104 | } |
| 105 | |
| 106 | // Compute left-over elements |
| 107 | for (; x < window_end_x; ++x) |
| 108 | { |
| 109 | const auto non_broadcast_v = *(non_broadcast_input_ptr + x); |
| 110 | auto res = |
| 111 | is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v; |
| 112 | if (is_broadcast_input_2) |
| 113 | { |
| 114 | res = static_cast<T>(-1) * res; |
| 115 | } |
| 116 | |
| 117 | *(output_ptr + x) = res; |
| 118 | } |
| 119 | }, |
| 120 | broadcast_input, non_broadcast_input, output); |
| 121 | } |
| 122 | else |
| 123 | { |
| 124 | // Clear X Dimension on execution window as we handle manually |
| 125 | input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 126 | input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 127 | |
| 128 | Iterator input1(src0, input1_win); |
| 129 | Iterator input2(src1, input2_win); |
| 130 | Iterator output(dst, win); |
| 131 | |
| 132 | execute_window_loop( |
| 133 | win, |
| 134 | [&](const Coordinates &) |
| 135 | { |
| 136 | const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr()); |
| 137 | const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr()); |
| 138 | const auto output_ptr = reinterpret_cast<T *>(output.ptr()); |
| 139 | |
| 140 | // Compute S elements per iteration |
| 141 | int x = window_start_x; |
| 142 | for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| 143 | { |
| 144 | const auto val1 = wrapper::vloadq(input1_ptr + x); |
| 145 | const auto val2 = wrapper::vloadq(input2_ptr + x); |
| 146 | const auto res = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2); |
| 147 | wrapper::vstore(output_ptr + x, res); |
| 148 | } |
| 149 | |
| 150 | // Compute left-over elements |
| 151 | for (; x < window_end_x; ++x) |
| 152 | { |
| 153 | const auto val1 = *(input1_ptr + x); |
| 154 | const auto val2 = *(input2_ptr + x); |
| 155 | *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2; |
| 156 | } |
| 157 | }, |
| 158 | input1, input2, output); |
| 159 | } |
| 160 | } |
| 161 | } // namespace cpu |
| 162 | } // namespace arm_compute |
| 163 | |
| 164 | #endif // ACL_SRC_CPU_KERNELS_SUB_NEON_IMPL_H |